Gut-VLM, 7 in 10 GPT-4o endoscopy descriptions had an error

* Derived: 1,232 mixed and 32 fully wrong descriptions out of 1,816 (Fig. 1 of the paper).

01What the experiment compares

Public image-text pairs are scarce in GI endoscopy; the authors cite one earlier set, Kvasir-VQA, with short answers. They build longer ones cheaply: GPT-4o describes each image, then gastroenterologists correct the description instead of writing one from nothing. The corrections leave a trace that ordinary datasets don’t have: which sentences the model got wrong.

A “report” here is a short paragraph about one still image, not the report of a procedure. Four open models (LLaVA-1.6-7B, DeepSeek-7B-VL, Qwen2-7B and mPLUG-Owl-2B, as the paper names them) are then tested three ways. The two fine-tuned versions share images and LoRA settings (rank 8, five epochs); they differ in the task, and with it in what the model is handed:

Table 1 — Two fine-tuned models with different inputs. Sources: Sections 2.3 and 3; Fig. 2 of the paper, which labels the second input “GPT-generated response”.

02What GPT-4o got wrong

So 1,264 descriptions, 69.6% derived, carried at least one tagged sentence. A mixed description has about 3.8 sentences, 1.4 of them tagged derived. The tags mark whole sentences, not the wrong words, a limit the authors name.

The paper’s Fig. 1 plots GPT-4o’s accuracy by aspect over the whole dataset: about 50% on the anatomical class (which part of the gut the image shows), against about 82% to 91% on colour, position, findings and polyp count, as measured on the chart our measurement. By that chart, GPT-4o got the anatomical class wrong about every other time.

The set is small and uneven: 1,450 images for training and 366 for testing, and 57% of the 1,816 show a normal landmark. Esophagitis, the smallest class, has 70 images, so 14 test images derived. All come from Kvasir, collected in one Norwegian health trust.

03What the models score

The paper’s main metric, QAAS, asks GPT-4o to pull 12 answers out of a generated description and counts how many match the 12 reference answers (Section 04).

1. Fine-tuning works. As released, LLaVA-1.6-7B scores 50.9% on QAAS; fine-tuned to write the description from the image, 83.1%; fine-tuned to correct, 90.9%. The other three models follow the same order.
2. The draft is a strong baseline that the paper lists but doesn’t compare against. GPT-4o’s unedited descriptions score 86.0%, which the authors call expected, partly because the references are corrected versions of them. All four writing models sit below that (82.9 to 83.7); all four correcting models sit above it (88.4 to 90.9).
3. The gain splits in two. Hallucination-aware fine-tuning beats standard fine-tuning by 5.0 to 7.8 points, the paper’s result. Measured from the draft instead, the correcting models add 2.4 to 4.9 points derived.

Table 2 — The same QAAS scores, measured from two starting points. QAAS (%) on the test set. “Gap”: correcting minus writing, the paper’s result. “Over the draft”: correcting minus GPT-4o’s unedited 86.0. Last column: that difference as a share of the draft’s 14.0 points of wrong answers, net of any new errors. Source: Table 1 of the paper; differences are ours and assume the reading in Section 01.

On that reading, a small open model trained on 1,450 examples removes, net, between 17% and 35% of the wrong answers in a GPT-4o draft derived: a useful checker on top of a large drafter. That is a different claim from “hallucination-aware beats standard fine-tuning”, which needs both models to start from the same input.

04Reading the two new metrics

QAAS, in the released scripts, takes two GPT-4o calls per description and one set comparison per answer. GPT-4o answers 12 questions from the text in one or two words (“N/A” if the text is silent); GPT-4o maps each answer to a category from a fixed list (for colour: red, pink, orange…; for the landmark: colon, cecum, pylorus, Z-line); an answer is correct when its categories overlap the reference’s. Staying silent where the reference is silent counts as correct.

R-Sim is one GPT-4o call: rate from 1 to 5 how well two descriptions match, with the 12 questions as a guide. Three cautions:

• The two metrics disagree about the draft. GPT-4o’s unedited descriptions get an R-Sim of 2.97, below every fine-tuned row, yet a higher QAAS than every writing model. And on DeepSeek, R-Sim goes down with hallucination-aware training, 3.76 to 3.63, while QAAS goes up.
• GPT-4o drafts, extracts and judges. A reviewer raised it; the authors agreed that several independent judges could be more robust. The paper reports that the extraction step itself errs 4.29% of the time on the reference texts, which were then fixed by hand; model outputs go through the same extractor unchecked, which the authors expect to affect all models alike.
• No intervals. One QAAS value rests on at most 366 × 12 = 4,392 answers. If answers were independent, a score near 88% would carry about ±1 point and a difference of two scores ±1.4 (unpaired; pairing on the shared images should tighten it); if the 12 answers of an image moved together, ±3.3 and ±4.7 derived. The 5.0 to 7.8 gaps clear both; of the four against the draft, only the largest clears the second.

One expert also rated 30 sampled LLaVA descriptions per version on a 1-to-5 scale: 1.90, 3.24 and 3.36. The 0.12 between the two fine-tuned versions is within the rater’s own repeat spread: the paper gives a coefficient of variation of 11.41%, about 0.37 per score at that level, so about ±0.19 on a difference between two 30-description averages from rater noise alone derived.

05What the paper leaves out

• The test-time input of the correcting model, and a control for it: a writing model given the same draft, or the correcting model given the standard model’s own output.
• Other ways to reduce hallucination. A reviewer asked for a comparison; the authors acknowledged the gap and deferred it to a journal version.
• Agreement between the four experts, and how many of them saw each description.
• The exact checkpoints. Model names are short, and two of the four citations are to papers on other models (the Qwen3 and DeepSeek-V3 language models).
• Training code and weights. Training ran on the ms-swift toolkit; the repository links no configuration file and no trained weights.
• The question-answering row. LLaVA fine-tuned to answer the 12 questions directly scores 87.9% on QAAS, above the writing version and the draft; the paper doesn’t set it against either.

06Takeaways

1. Score the draft before you score the corrector. Here the draft is worth 86.0% on the paper’s own metric; a corrector’s 90.9% is +4.9 over it, not +40 over the model as released.
2. Keep the corrections when experts fix LLM drafts. Sentence-level tags cost little extra and turn a caption set into training data for a checker. Expect real work: 7 drafts in 10 needed a change.
3. Check the anatomical site separately. It is where the draft was weakest, about one in two, and where correcting did no better than writing.
4. Don’t let the drafter be the only judge. Two GPT-4o-based scores rank GPT-4o’s own text in opposite ways.

07What you can use today

• The annotations. In dataset/Gut-VLM of the repository: all_annotations.json (2.2 MB; draft, sentence tags and corrected text for the complete dataset, per the README), VQA_format_all_annotations.json (3.0 MB; the 12 answers), and train_test_split/ with train.json, test.json, question-answer versions and one file per class. training_style/hallucinated_aware_train.json (3.0 MB) is described as the detect-and-correct training file. We read the listing and the README, not the files.
• The images are not in the repository. Download kvasir-v2.zip (2.3 GB) from Simula’s Kvasir page.
• The licence. The repository has no licence file, so the annotations come with no stated reuse terms: ask the authors before redistributing. For the images, Simula’s page restricts Kvasir to research and education and forbids commercial use without written permission.
• The evaluation scripts. eval_scripts/ holds the QAAS chain (caption2vqa.py, parse_vqa.py, evaluate_vqa.py) and R-Sim (evaluate_caption.py). They call gpt-4o through the OpenAI API, so scores can drift as that model changes. The prompts and synonym lists are in the scripts.
• The rule. When a model’s job is to fix another model’s text, report three numbers: the text before, the text after, and what a model without the draft achieves.

ASources

1. Khanal, B., Pokhrel, S., Bhandari, S., Rana, R., Shrestha, N., Gurung, R.B., Linte, C., Watson, A., Shrestha, Y.R., Bhattarai, B. Hallucination-Aware Multimodal Benchmark for Gastrointestinal Image Analysis with Large Vision-Language Models. MICCAI 2025, LNCS 15969, pp. 235–245: paper page with reviews and rebuttal · PDF · DOI. The reference for every number here.
2. The earlier version, arXiv:2505.07001 v2 (22 Jun 2025): same result table and counts; it lacks the per-class image counts and the longer metric descriptions of the MICCAI version.
3. The repository, read on 13 Oct 2026 through the GitHub API: file listing, README, the four evaluation scripts and the two sample files in results_example/ (one model’s test outputs and the reference texts). No annotation file (dataset/, training_style/) was opened.
4. Pogorelov, K., et al. Kvasir: a multi-class image dataset for computer aided gastrointestinal disease detection. ACM MMSys 2017; dataset page for the download and the terms of use.